Useful Nonrobust features are ubiquitous in biomedical images
Coenraad Mouton ⋆† Randle Rabe † Niklas C Koser ⋆ Nicolai Krekiehn ⋆ Christopher Hansen ⋆ Jan-Bernd Hövener ⋆ Claus-C. Glüer ⋆ · 2026
Read Useful Nonrobust features are ubiquitous in biomedical images by Coenraad Mouton ⋆† Randle Rabe † Niklas C Koser ⋆ Nicolai Krekiehn ⋆ Christopher Hansen ⋆ Jan-Bernd Hövener ⋆ Claus-C. Glüer ⋆ (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-22579.txt.
Description
Mouton and six colleagues ask whether medical-image networks learn useful nonrobust features, telling patterns humans cannot read and hostile tweaks can break. Models trained only on those features score above chance across five MedMNIST sorting tasks, confirming their value in training data. Hardened models leaning on robust features give up some in-distribution accuracy but perform much better under data shifts on MedMNIST-C. The upshot is a trade-off: those features lift accuracy yet harm out-of-distribution performance, so the balance should follow the use setting.
Catalog record
| CREATOR | Coenraad Mouton ⋆† Randle Rabe † Niklas C Koser ⋆ Nicolai Krekiehn ⋆ Christopher Hansen ⋆ Jan-Bernd Hövener ⋆ Claus-C. Glüer ⋆ |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 19686 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2604.22579